Which reported effects track which receptor profiles? Asked across a whole corpus of first-person reports rather than one at a time, with every extracted fact traceable to the quotation behind it.
An individual report is an anecdote. Aggregated across many reports and joined to each molecule's receptor-binding profile, the same material becomes analysable.
PsycheGraph represents each experience as a reified event, linked to the substances taken, to the effects described (each with an intensity, a time of onset, and a verbatim supporting quotation), and to any lasting or adverse outcomes. A parallel molecular layer records structure and receptor affinities, so reported phenomenology can be examined against receptor pharmacology rather than discussed in isolation.
The current corpus is drawn from Erowid as a proof-of-concept basis. The schema is not specific to it.
SMILES and InChI, chemical class and prodrug relationships, so a candidate sits among its structural neighbours (PubChem).
Binding affinities (Kᵢ) across serotonin, dopamine, NMDA and opioid targets — the axis a candidate's profile is compared on (PDSP).
Effects recorded as timed, quoted events within a hierarchical taxonomy, so reported experience can be queried rather than only read.
Self-reported lasting effects and adverse events, surfaced as naturalistic signal rather than case-by-case anecdote.
An early-stage prototype. The figures below are read from the running graph at page load.
Corpus figures withheld
Report, substance and word counts that would draw on Erowid experience reports are withheld pending a data-rights agreement. The enrichment coverage below is measured against PubChem and PDSP, which carry no such restriction.
Illustrative rather than a result. The intent is to show the form of the query.
-- Which phenomena covary with high 5-HT2A affinity (Kᵢ < 100 nM)? phenomenon mean_intensity n_substances confidence transpersonal.unity x n CI, low-n flagged perceptual.visual.geometry x n CI, low-n flagged -- joined: substance -> receptor_binding -> receptors -- substance <- experience event -> phenomenon
The evidence is self-reported, and the design treats that as a first-order constraint.
Experience reports carry well-documented biases. Reporting is subject to selection effects; drug identity and dose are typically unverified; polydrug use confounds attribution; timing is inconsistent; vocabulary is culturally primed; and outcomes are self-assessed rather than clinically measured. The response is statistical: per-statistic sample counts, confidence intervals, low-sample flags, and complete provenance, with every extracted fact linked to the quotation from which it was derived.
A second limitation sits above the corpus itself. Phenomenology and outcome
data of this kind cannot be read off free-text reports reliably: the known
failure modes are most serious on adverse events, dose, and subevent timing, and
nothing in this graph's phenomenology or outcome layer should be read as
complete or as verified on those points. Receptor pharmacology (PDSP) and
molecular structure (PubChem) are separately sourced and not subject to that
limitation. Nothing on this site is dosing guidance, a safety assurance, or
medical advice — see docs/methodology-and-limitations.md in the
repository for the full methodology and known limitations.
The approach has precedent. Ballentine, Friedman and Bzdok (Science Advances, 2022) applied natural-language processing to a corpus of Erowid reports across 40 receptor subtypes, and Zamberlan et al. (2018) reported that binding-affinity similarity tracks the semantic similarity of reports. PsycheGraph builds on that work as a reusable, provenance-tracked graph in which the molecular join is first-class.
The example corpus of first-person reports for this proof of concept. Attribution is preserved.
Molecular structure: SMILES, InChI, formula, and weight.
Receptor binding affinities (UNC and NIMH PDSP).
A research programme runs behind the graph; results are published, not pre-announced.
An active research programme connects the graph to mechanistic models of psychedelic action.
PsycheGraph is an independent research project. Critical feedback is welcome, and I would be glad to hear from anyone who regards the method as useful or as flawed: discovery teams weighing candidates against reported phenomenology, anyone working on naturalistic safety signal, researchers who want to inspect the provenance and the caveats, and anyone with phenomenological or pharmacological data they would like to contribute.
Share your details if you would like to discuss the research, a collaboration, or data you would like to contribute.